Re: [PATCH] stackprotector: fix multi-word cross-builds

[Tejun Heo <tj@kernel.org>]

Ingo Molnar wrote:
> * Ingo Molnar <mingo@elte.hu> wrote:
> 
>> * Tejun Heo <tj@kernel.org> wrote:
>>
>>> Ingo Molnar wrote:
>>>>>> I'll try other compilers but which version are you using?  The
>>>>>> difference is that before the patchset, -fno-stack-protector was
>>>>>> always added whether stackprotector was enabled or not so this problem
>>>>>> wasn't visible (at the cost of bogus stackprotector of course).  We'll
>>>>>> probably need to add __stack_chk_guard or disable if gcc generates
>>>>>> such symbol.  I'll play with different gccs.
>>>>> Can't reproduce with gcc-4.1 or 4.2.  Any chance you're using distcc
>>>>> w/ a build machine w/ glibc < 2.4?  __stack_chk_guard is the symbol
>>>>> gcc fetches stack canary from if TLS is not supported, so somehow gcc
>>>>> thought that TLS wasn't available while building head64.
>>>> yeah - i also used distcc. Maybe the nostackp makefile magic gets confused
>>>> about that?
>>> It seems that even with the same gcc versions, gcc built against libc
>>> w/o TLS support generates __stack_chk_guard, so if you mix the two
>>> flavors, the has-stack-protector check can be compiled on machines w/
>>> TLS while some other files end up being built on machines w/o TLS
>>> support thus circumventing the support check.  Can you please see
>>> whether non-distcc build fails too?
>> That build succeeds:
>>
>> rhea:~/tip> make -j30 bzImage ARCH=x86_64 CROSS_COMPILE='/opt/crosstool/gcc-4.2.3-glibc-2.3.6/x86_64-unknown-linux-gnu/bin/x86_64-unknown-linux-gnu-'
>> /home/mingo/tip/arch/x86/Makefile:82: stack protector enabled but no compiler support
>>   CHK     include/linux/version.h
>> [...]
>> BFD: arch/x86/boot/compressed/vmlinux.bin: warning: allocated section `.bss' not in segment
>> [...]
>> Root device is (8, 3)
>> Setup is 11996 bytes (padded to 12288 bytes).
>> System is 5690 kB
>> CRC be1b2e21
>> Kernel: arch/x86/boot/bzImage is ready  (#3)
>>
>> Some shell variable expansion bug? If CROSS_COMPILE is not a single word
>> we fail to detect the compiler borkage at arch/x86/Makefile line 82?
> 
> Yep - i'm testing the fix below now - it's looking good so far.

Ah... okay.  No wonder I couldn't reproduce the problem.  :-)

Thanks for hunting it down.

-- 
tejun